Why Using AI to Write Substantive Text Is Usually a Bad Idea

TL;DR – AI‑generated prose is harmful

Using an LLM to produce a full, substantive piece of writing is generally a mistake because it bypasses the thinking that writing provides, it injects vague or inaccurate statements that are hard to spot, and it breaks the implicit contract that the reader expects the author’s own ideas.

1. Writing is a thinking tool

"Writing about something, even something you know well, usually shows you that you didn’t know it as well as you thought." – Paul Graham

The act of translating thoughts into words forces you to clarify arguments, expose contradictions, and fill gaps in evidence.
Clara Collier (Complex Systems Podcast) observes that every sentence revision reveals whether an outline is sound; an LLM cannot flag a bad outline because it never thinks about the structure.
Patrick McKenzie adds that forcing a multi‑page argument into prose surfaces holes that would otherwise remain hidden.

Conclusion: If you skip the writing step, you skip the mental work that makes the final argument reliable.

2. AI prose is densely vague and subtly wrong

Example: Claude Fable 5’s paragraph on AI‑chip smuggling contains numerous imprecise claims (e.g., “export controls are only as strong as enforcement”, “chips are compact”, “sprawling global supply chains make smuggling attractive”).
Each phrase is either a truism, factually inaccurate, or lacking context—issues that a knowledgeable author would notice and correct.
Eric Schwitzgebel notes that human experts choose words that reflect nuanced sensitivities; LLMs lack this fine‑grained awareness and produce “hallucinated” statements that sound plausible.
The author’s own experience: AI‑generated meeting‑note summaries still introduced subtle vagueness despite detailed prompts and full transcripts.

Conclusion: AI‑written text is riddled with low‑visibility errors that erode credibility and require expert knowledge to detect.

3. Undisclosed AI writing is rude and misleading

Readers assume a text reflects the author’s own reasoning; an AI‑generated draft violates that implicit contract.
When the author discovers AI output in a colleague’s email or report, trust drops instantly.
Labeling AI‑generated passages can mitigate the problem, but most people avoid disclosure, turning the practice into a form of intellectual dishonesty.

Conclusion: Presenting AI‑written prose as your own work is ethically questionable and damages the author‑reader relationship.

4. When (if ever) is AI‑assisted writing acceptable?

Use case	Recommended AI role
Transcribing audio, data analysis, literature search, brainstorming	Full generation (no disclosure needed)
Line‑level copy‑editing, tightening prose, fixing grammar	Human‑approved edit; each change must be consciously accepted
Short, formulaic communications (e.g., routine status emails)	Draft generation followed by brief human edit
Citing AI output directly (e.g., “Claude says…”)	Allowed if the attribution is clear
Full‑text generation for a public audience	Not recommended, even with a disclaimer; quality and trust suffer

Conclusion: AI is valuable as a tool for peripheral tasks, not as a substitute for the core writing process.

5. Community perspectives from Hacker News

"Reviewing and improving an LLM‑generated work document was slower and produced lower quality than writing it from scratch." – @patrickmay

Several commenters echo the main points: AI‑generated prose is often vague, requires extra effort to clean up, and can erode personal style.
A minority argue that LLMs can improve technical writing by offering alternative phrasings and catching filler words (e.g., @tptacek, @atdt), but they stress that the author must retain final control.
Others note that for non‑native speakers, AI can help with translation, yet still recommend clear attribution (e.g., @diego_moita).
A few dismiss the warning as “Luddism” or claim the productivity gains outweigh quality loss (e.g., @numitus), but these views are in the clear minority among the top‑voted comments.

Conclusion: The consensus on HN aligns with the author’s thesis: AI should augment, not replace, substantive writing.

6. Practical takeaways

Never let an LLM write a draft you intend to publish without heavy human revision. The mental work of drafting is essential for sound reasoning.
Use AI for peripheral tasks only – transcription, data summarization, idea brainstorming, and deterministic copy‑editing.
Always disclose AI‑generated excerpts when you include them verbatim; treat them as quoted sources, not as your own prose.
Treat the LLM as a critic, not a co‑author. Prompt it to point out ambiguities, suggest alternatives, or summarize, then decide yourself what to keep.
Invest in your own writing muscles. Iterative drafting, even if slower, yields clearer arguments and preserves trust with readers.

7. Final thought

Even as LLMs become more capable, the process of writing remains a uniquely human exercise that shapes thought, ensures precision, and upholds the ethical contract between author and audience. Until models can genuinely think and self‑audit, delegating substantive writing to AI is, in most cases, a shortcut that costs more than it saves.